Feasibility of AI-assisted CBCT-guided Stereotactic Online Adaptive Radiotherapy (CT-STAR) for Upper Abdominal SBRT: Results of a Prospective in Silico Clinical Trial

Abstract

Stereotactic online adaptive radiotherapy (STAR) has been shown to improve the dosimetric therapeutic index (DTI) of SBRT for pancreatic cancer and other upper abdominal malignancies. The feasibility of this technique using a commercially available, online adaptive platform coupled with a ring-gantry Linac and artificial intelligence (AI)-assisted workflows, using cone-beam computed tomography (CBCT) image-guidance, has yet to be evaluated. We conducted a prospective in silico imaging clinical trial of CBCT-guided STAR (CT-STAR), with the hypothesis that CT-STAR would be feasible in silico and would improve the DTI for upper abdominal SBRT. Five patients with upper abdominal malignancies (3 pancreatic, 1 liver, 1 oligometastatic lymph node) who were otherwise undergoing definitive SBRT in 5 fractions (fx) were imaged using high-quality daily CBCT on the ring gantry Linac on each of their clinical treatment days. For all patients, initial plans (50Gy/5fx) were created using clinical planning CT images. CT-STAR was then simulated using the daily CBCT images, comprising all steps of the CT-STAR workflow: image approval, AI-driven autocontouring of influential organs-at-risk (OARs; liver, duodenum, stomach), physician editing of auto-contours, physician editing of the clinical target volume (CTV) and additional OARs (bowel), automated adaptive plan generation and dose volume histogram (DVH) comparison to the initial plan as projected on the daily anatomy, plan approval, pre-delivery quality assurance, and plan delivery. All plans used a strict isotoxicity approach, such that OAR constraints were met at each fx, with PTV coverage increased or decreased as permitted. Feasibility was defined as successful completion of all steps of the CT-STAR process in > 80% of simulated fx. A total of 23 CT-STAR fx were simulated. Median CTV and PTV at baseline were 86.9 cm3 (range, 5.0-115.8 cm3) and 170.9 cm3 (19.4-216.2 cm3). 61 AI OAR contours were autogenerated and reviewed with editing by the physician. AI driven auto-planning led to creation of acceptable plans (OAR constraints met, coverage acceptable) for all fx. In 100% of fx, the adaptive plan was selected over the initial plan, either because the baseline plan initially violated >/ = 1 OAR constraints (15/23fx, 65.2%) or because CTV/PTV coverage was improved by >5% (institutional clinical threshold for adaptation). After all plan re-optimization, the median per-fraction value for the mean PTV dose was 10.61 Gy (9.88-11.33 Gy) and the median max PTV dose was 13.68 Gy (12.51-14.98 Gy). The most common OAR constraint violations prior to adaptation were the duodenum (65.2% fx) and stomach (52.2% fx). 100% of violations were resolved with adaptation. All steps of CT-STAR were successfully completed in 23/23 fx. AI-assisted, CT-STAR is feasible in silico and improves the DTI of SBRT to upper abdominal cancers. Prospective clinical evaluation of this approach is indicated.